Telemetry

Honeybee’s telemetry system captures operational metrics from every layer of the stack. It’s local-first — everything writes to disk by default. Cloud sync is opt-in and sends only aggregated summaries, never raw content.

12 event types

Event	Source	What it captures
`llm_call`	Runner	Model, tokens (prompt/completion), latency, cost
`llm_error`	Runner	Error type, model, retry count
`tool_call`	Runner	Tool name, duration, success/failure
`agent_complete`	Runner	Exit reason (cost_limit, max_iterations, halt, error, done)
`context_compaction`	Runner	Tokens before/after, compaction trigger
`guard_scan`	Guard	Side (write/read/snapshot), score, action, findings count
`agent_spawn`	Orchestrator	Agent ID, role, type, model
`agent_exit`	Orchestrator	Agent ID, exit reason, duration
`agent_kill`	Orchestrator	Agent ID, kill reason
`protocol_start`	Orchestrator	Protocol name, agent count
`protocol_end`	Orchestrator	Duration, total tokens, total cost
`dance_call`	WebSocket	Tool name, caller role, latency

Local-first storage

Telemetry always writes to local JSONL files:

~/.honeyb/projects/<slug>/telemetry/2026-02-15.jsonl

Each line is a JSON event with timestamp, type, and payload. Read with any tool:

# View with wgl
wgl telemetry overview --local
wgl telemetry events --local --type llm_call
wgl telemetry timeline --local

# Or just read the file
cat ~/.honeyb/projects/myapp/telemetry/2026-02-15.jsonl | jq .

Cloud sync (opt-in)

To enable cloud telemetry, set two environment variables:

export TELEMETRY_ENDPOINT=https://carapace.honeyb.dev/v1/telemetry
export TELEMETRY_API_KEY=your-key

Cloud sync sends 5-minute aggregated summaries, not raw events:

Total tokens (prompt + completion)
Event counts by type
Average latency
Error rates
Guard scan summary (pass/warn/block counts)

Never sent: Actual prompts, responses, tool arguments, file contents, or any user data.

Dashboard

The incubator dashboard includes a Metrics tab (7th tab) with live telemetry:

Stat cards: Total tokens, total cost, agent count, error rate
Event breakdown: Bar chart of events by type
Numeric averages: Latency, tokens per call, cost per agent

Metrics push via WebSocket every 10 seconds using getSnapshot().

CLI commands

# Aggregate overview (counts, totals, averages)
wgl telemetry overview

# Recent events (filterable by type)
wgl telemetry events --type guard_scan --limit 20

# Daily breakdown by event type
wgl telemetry timeline

# Force local-only (skip cloud even if configured)
wgl telemetry overview --local

Privacy guarantees

Structural metadata only: Names, scores, counts, latency, booleans
Never captures: Prompts, responses, tool arguments, file contents
Local by default: No data leaves your machine without explicit configuration
Aggregated sync: Cloud receives summaries, not individual events
Open source: The telemetry reporter code is in @honeybee-ai/hivemind-sdk/telemetry — read it yourself